Getting started

How to create an AI agent around your expertise

A practical starting point for an AI agent with your methods, tools, and voice. Learn what to prepare, what Orage provides, and what still needs setup.

Start with one person and one useful job

A good first agent has a clear purpose. A team mentor might help a manager prepare for a difficult conversation. A garden guide might help someone choose plants for a small balcony. Start with the moment someone would reach for it, then write down what a useful result looks like.

Choose a narrow first job you can judge yourself. This makes it easier to tell whether the agent applies your method, asks the right questions, and knows when it should stop.

Write down how you think

Prepare the instructions that make the agent yours: what it helps with, how it speaks, which questions it asks first, and the boundaries it must respect. Add repeatable methods as skills. In Orage, instructions and skills are Markdown files you can edit in the file workspace.

For a team mentor, a method might be: ask for a concrete example, separate facts from assumptions, then propose one next conversation. A few clear steps give you something you can test and improve.

Run an instance and choose a model

An Orage instance holds your agent's settings, files, tools, conversations, and access rules. Follow the repository's current setup instructions to run one. You need a compatible host and model-provider access; Orage is not an instant, no-code hosted signup.

Once the instance is running, its owner can open Admin, choose a model, and edit the agent's name, instructions, and skills. Keep the first version small enough to review end to end.

Connect only the tools the job needs

Some agents only need a useful conversation. Others need to read files, look things up, or take an action. Give the agent the tools it needs for the first job and review the permissions attached to each one.

A connection being available does not mean every action should be allowed. Keep credentials private, check the information an action will send, and retain confirmation where the consequences require it.

Test the awkward cases, then choose a channel

Try a clear request, an incomplete one, a request outside the agent's scope, and a failed tool call. Check the reply and what the agent actually did. Reopen the conversation and confirm that the result still makes sense.

Start with the web app. When the experience is useful, configure a messaging channel if that better fits your audience. WhatsApp, iMessage, and Telegram have different provider and account requirements. Verify the real first-message path before inviting people.